{
  "schemaVersion": "dxrg-trading-agent-prompt-compilation-ablation-registry/v1",
  "version": "1.0.0",
  "title": "DXRG Trading-Agent Prompt-Compilation Ablation Registry",
  "description": "Eight unrun comparison templates for testing prompt-compilation changes while freezing the model, state, memory, action schema, policy, execution adapter, and evidence class.",
  "publishedDate": "2026-07-31",
  "publisher": "DX Research Group",
  "canonicalPage": "https://www.dxrg.ai/blogs/trading-agent-mandate-compiler",
  "methodology": "Register one compiler change before evaluation. Freeze every named component, run matched mandate and state fixtures, score the typed action or abstention, and preserve the path through policy, execution, settlement, and reconciliation. Result fields remain null until a completed evaluation supplies evidence.",
  "evidenceBoundary": "This registry is an evaluation protocol with unrun result fields. It does not establish that prompt compilation improves trading skill, alpha, profitability, universal safety, or transfer beyond a completed registered comparison.",
  "defaultFrozenComponents": [
    "model and model version",
    "sampling parameters",
    "authenticated mandate fixtures",
    "market and portfolio state snapshots",
    "memory snapshot and provenance",
    "typed action schema",
    "deterministic policy version",
    "execution adapter and venue rules",
    "evaluation window and evidence class"
  ],
  "requiredOutcomeFields": [
    "compiled mandate diff",
    "typed action or abstention",
    "policy result",
    "final payload status",
    "venue acknowledgement",
    "settlement result",
    "reconciled portfolio effect"
  ],
  "sources": [
    {
      "id": "dx-terminal-pro-paper",
      "title": "Operating-Layer Controls for Onchain Language-Model Agents Under Real Capital",
      "url": "https://arxiv.org/abs/2604.26091",
      "version": "v1",
      "locator": "Abstract; Control-Loop Results; Table 4; Trace Reuse",
      "use": "First-party source for compilation-sensitive failure cases and the linked mandate-to-settlement trace.",
      "boundary": "One bounded deployment and its controlled pre-launch tests; reliability results are separate from trading skill and profitability."
    },
    {
      "id": "atlas-adaptive-opro",
      "title": "ATLAS: Adaptive Trading with LLM AgentS Through Dynamic Prompt Optimization and Multi-Agent Coordination",
      "url": "https://arxiv.org/abs/2510.15949",
      "version": "v5",
      "locator": "Sections 3, 5, 6.1, and 6.4; Appendix F",
      "use": "Trading-specific primary research that compares fixed prompts, reflection, and dynamic prompt optimization inside its own order-aware framework.",
      "boundary": "A separate equity framework, market design, prompt optimizer, and result set; it does not validate DXRG methods or transfer to another runtime."
    },
    {
      "id": "prompt-optimization-coin-flip",
      "title": "Prompt Optimization Is a Coin Flip: Diagnosing When It Helps in Compound AI Systems",
      "url": "https://arxiv.org/abs/2604.14585",
      "version": "v2",
      "locator": "Sections 3, 4, 6, and 7",
      "use": "Primary research on component interaction, zero-shot baselines, repeated optimization runs, and headroom checks before optimization.",
      "boundary": "General compound-AI tasks rather than trading agents; the reported thresholds and costs are specific to the paper's setup."
    },
    {
      "id": "promptagent",
      "title": "PromptAgent: Strategic Planning with Language Models Enables Expert-level Prompt Optimization",
      "url": "https://arxiv.org/abs/2310.16427",
      "version": "v2",
      "locator": "Sections 3.1, 3.2, 4.1, and 4.2",
      "use": "Primary research showing an explicit search process over prompt candidates with error feedback and held task evaluations.",
      "boundary": "Non-trading language tasks; prompt-search results do not establish mandate fidelity, execution validity, or market performance."
    }
  ],
  "claimPlan": [
    {
      "claimId": "compiler-change-requires-matched-freeze",
      "status": "SOURCE_PLAN_ONLY",
      "eligibleWording": "A prompt-compilation comparison should register the changed compiler component and freeze the surrounding evaluation subject.",
      "sourceIds": [
        "dx-terminal-pro-paper",
        "prompt-optimization-coin-flip"
      ],
      "boundary": "DXRG evaluation recommendation informed by separate first-party and third-party methods."
    },
    {
      "claimId": "prompt-optimization-can-vary-by-setup",
      "status": "SOURCE_PLAN_ONLY",
      "eligibleWording": "Published prompt-optimization results vary by task, model, framework, and evaluation design.",
      "sourceIds": [
        "atlas-adaptive-opro",
        "prompt-optimization-coin-flip",
        "promptagent"
      ],
      "boundary": "Cross-paper methodological observation, not a pooled effect estimate or DXRG result."
    },
    {
      "claimId": "trading-prompt-optimization-exists",
      "status": "SOURCE_PLAN_ONLY",
      "eligibleWording": "ATLAS evaluates a dynamic prompt-optimization method inside its own order-aware equity trading framework.",
      "sourceIds": [
        "atlas-adaptive-opro"
      ],
      "boundary": "Describes the cited paper only."
    },
    {
      "claimId": "search-based-prompt-optimization-exists",
      "status": "SOURCE_PLAN_ONLY",
      "eligibleWording": "PromptAgent frames prompt optimization as search with error feedback and evaluates it on non-trading language tasks.",
      "sourceIds": [
        "promptagent"
      ],
      "boundary": "Describes the cited paper only."
    }
  ],
  "ablations": [
    {
      "ablationId": "precedence-order",
      "component": "instruction precedence",
      "changeUnderTest": "Change only the order and explicit source labels used to resolve conflicting natural-language and structured instructions.",
      "fixtureClass": "contradictory mandate",
      "primaryMeasure": "share of typed actions and abstentions that follow the registered winning instruction",
      "requiredTraceFields": [
        "source map",
        "precedence decision",
        "compiled mandate hash",
        "typed action or abstention",
        "policy result"
      ],
      "validityChecks": [
        "losing instruction remains auditable",
        "no mixed-version fields",
        "same state snapshot for both arms"
      ],
      "result": {
        "state": "UNRUN",
        "value": null,
        "sampleSize": null,
        "confidenceInterval": null
      },
      "verdict": null,
      "sourceIds": [
        "dx-terminal-pro-paper"
      ],
      "boundary": "Mandate-fidelity result only."
    },
    {
      "ablationId": "critical-fact-placement",
      "component": "rendering order",
      "changeUnderTest": "Move one unchanged market mechanic or cost field while preserving its wording and source.",
      "fixtureClass": "critical fact salience",
      "primaryMeasure": "use of the fixed fact in the typed action decision and its linked reasoning trace",
      "requiredTraceFields": [
        "rendered prompt diff",
        "fact source",
        "typed action or abstention",
        "policy result",
        "settlement result"
      ],
      "validityChecks": [
        "fact value and wording unchanged",
        "matched invocation fixtures",
        "behavioral measure separated from return"
      ],
      "result": {
        "state": "UNRUN",
        "value": null,
        "sampleSize": null,
        "confidenceInterval": null
      },
      "verdict": null,
      "sourceIds": [
        "dx-terminal-pro-paper"
      ],
      "boundary": "Fact-use and action-behavior result rather than profitability."
    },
    {
      "ablationId": "structured-control-rendering",
      "component": "structured settings",
      "changeUnderTest": "Change only how one authenticated typed setting is rendered into model context.",
      "fixtureClass": "setting gradient",
      "primaryMeasure": "monotonic response of typed actions or abstentions across registered setting values",
      "requiredTraceFields": [
        "setting source and unit",
        "compiled value",
        "typed action or abstention",
        "policy result",
        "reconciled portfolio effect"
      ],
      "validityChecks": [
        "same setting values and units",
        "policy limits unchanged",
        "gradient separated from action quality"
      ],
      "result": {
        "state": "UNRUN",
        "value": null,
        "sampleSize": null,
        "confidenceInterval": null
      },
      "verdict": null,
      "sourceIds": [
        "dx-terminal-pro-paper"
      ],
      "boundary": "Behavioral responsiveness rather than evidence that a higher setting is better."
    },
    {
      "ablationId": "historical-context-authority",
      "component": "memory authority labels",
      "changeUnderTest": "Change only the label that distinguishes prior reasoning from current mandate authority.",
      "fixtureClass": "stale context conflict",
      "primaryMeasure": "rate at which historical context is promoted into an active rule",
      "requiredTraceFields": [
        "memory provenance",
        "effective mandate version",
        "rendered label",
        "typed action or abstention",
        "policy result"
      ],
      "validityChecks": [
        "memory contents unchanged",
        "current mandate unchanged",
        "compound changes excluded"
      ],
      "result": {
        "state": "UNRUN",
        "value": null,
        "sampleSize": null,
        "confidenceInterval": null
      },
      "verdict": null,
      "sourceIds": [
        "dx-terminal-pro-paper"
      ],
      "boundary": "Authority-label test; it cannot inherit the compound 57% to 3% result."
    },
    {
      "ablationId": "abstention-path",
      "component": "typed abstention instruction",
      "changeUnderTest": "Change only the compiled wording and placement of the available abstention action.",
      "fixtureClass": "insufficient or conflicting state",
      "primaryMeasure": "valid abstention rate on fixtures where new exposure is ineligible",
      "requiredTraceFields": [
        "state eligibility result",
        "abstention schema",
        "typed action or abstention",
        "policy result",
        "recovery state"
      ],
      "validityChecks": [
        "same abstention schema",
        "same ineligible fixtures",
        "missed trades excluded from success measure"
      ],
      "result": {
        "state": "UNRUN",
        "value": null,
        "sampleSize": null,
        "confidenceInterval": null
      },
      "verdict": null,
      "sourceIds": [
        "dx-terminal-pro-paper"
      ],
      "boundary": "Correct withholding of action rather than a return result."
    },
    {
      "ablationId": "action-schema-instructions",
      "component": "action contract rendering",
      "changeUnderTest": "Change only the compiled representation of the fixed typed action schema.",
      "fixtureClass": "schema boundary",
      "primaryMeasure": "rate of schema-valid proposed actions before deterministic policy",
      "requiredTraceFields": [
        "schema version",
        "rendered action contract",
        "model output",
        "parsed typed action",
        "policy result"
      ],
      "validityChecks": [
        "schema and parser unchanged",
        "same model and sampling settings",
        "syntax separated from mandate fidelity"
      ],
      "result": {
        "state": "UNRUN",
        "value": null,
        "sampleSize": null,
        "confidenceInterval": null
      },
      "verdict": null,
      "sourceIds": [
        "atlas-adaptive-opro",
        "promptagent"
      ],
      "boundary": "Construction validity rather than economic quality."
    },
    {
      "ablationId": "mandate-version-update",
      "component": "effective-version rendering",
      "changeUnderTest": "Change only how a newer authenticated mandate version supersedes an in-flight compiled version.",
      "fixtureClass": "mid-inference mandate update",
      "primaryMeasure": "rate of submissions correctly canceled, recompiled, or withheld after authority changes",
      "requiredTraceFields": [
        "old and new mandate versions",
        "effective times",
        "compiled mandate hash",
        "policy result",
        "final payload status"
      ],
      "validityChecks": [
        "same update timing fixtures",
        "no cross-version field merge",
        "submission status proven"
      ],
      "result": {
        "state": "UNRUN",
        "value": null,
        "sampleSize": null,
        "confidenceInterval": null
      },
      "verdict": null,
      "sourceIds": [
        "dx-terminal-pro-paper"
      ],
      "boundary": "Authority-version integrity rather than trading skill."
    },
    {
      "ablationId": "final-payload-binding",
      "component": "compiler-to-policy binding",
      "changeUnderTest": "Change only the compiler identifier and field bindings carried into final venue-ready payload validation.",
      "fixtureClass": "normalization and retry",
      "primaryMeasure": "rate of final payloads proven to match the active mandate and state versions",
      "requiredTraceFields": [
        "compiled mandate hash",
        "state snapshot identifier",
        "proposed action",
        "final payload",
        "venue acknowledgement",
        "reconciled outcome"
      ],
      "validityChecks": [
        "same normalization and retry fixtures",
        "policy version unchanged",
        "unknown acknowledgements remain unresolved"
      ],
      "result": {
        "state": "UNRUN",
        "value": null,
        "sampleSize": null,
        "confidenceInterval": null
      },
      "verdict": null,
      "sourceIds": [
        "dx-terminal-pro-paper",
        "prompt-optimization-coin-flip"
      ],
      "boundary": "Payload-binding and reconciliation result rather than profitability or transfer."
    }
  ]
}